Goodhart's Village: Using LLM-Mafia to Study Deception

The social deduction game Mafia centres on reasoning under information asymmetry, where an informed minority must mislead an uninformed majority, making it a useful setting for studying deception in large language models (LLMs). Although LLMs have seen rapid progress in areas such as reasoning and language understanding, their ability to engage in social reasoning under uncertainty remains poorly understood. In this work, we study deceptive behaviour in a six-player implementation of the full Mafia game, extending prior work based on a simplified variant. By varying behavioural instructions from strict honesty to a “win at all costs” objective, we examine how explicit prompting interacts with the structural demands of adversarial roles. Comparing agents’ private reasoning with their public statements, we find that Mafia agents display consistently high levels of deception even when instructed not to lie, while cooperative roles adapt their behaviour more flexibly in response to perceived threat. Overall, the results suggest that role structure and game incentives dominate behavioural prompting, supporting Mafia as a useful benchmark for analysing deception and social reasoning in LLMs.